model-net-mpi-replay.c 40.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/*
 * Copyright (C) 2014 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */
#include <ross.h>
#include <inttypes.h>

#include "codes/codes-workload.h"
#include "codes/codes.h"
#include "codes/configuration.h"
#include "codes/codes_mapping.h"
#include "codes/model-net.h"
#include "codes/rc-stack.h"
#include "codes/quicklist.h"

17 18
/* turning on track lp will generate a lot of output messages */
#define TRACK_LP -1
19
#define TRACE -1
20
#define MAX_WAIT_REQS 512
21
#define DBG_MPI_SIM 0
22 23 24 25 26 27 28 29 30 31 32 33 34 35 36

char workload_type[128];
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;

/* Doing LP IO*/
static char lp_io_dir[256] = {'\0'};
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
37
typedef int32_t dumpi_req_id;
38 39 40 41 42

static int net_id = 0;
static float noise = 5.0;
static int num_net_lps, num_nw_lps;

43 44 45 46
#define CS_LP_DBG 0
#define lprintf(_fmt, ...) \
        do {if (CS_LP_DBG) printf(_fmt, __VA_ARGS__);} while (0)

47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

double max_time = 0,  max_comm_time = 0, max_wait_time = 0, max_send_time = 0, max_recv_time = 0;
double avg_time = 0, avg_comm_time = 0, avg_wait_time = 0, avg_send_time = 0, avg_recv_time = 0;

/* global variables for codes mapping */
static char lp_group_name[MAX_NAME_LENGTH], lp_type_name[MAX_NAME_LENGTH], annotation[MAX_NAME_LENGTH];
static int mapping_grp_id, mapping_type_id, mapping_rep_id, mapping_offset;

/* runtime option for disabling computation time simulation */
static int disable_delay = 0;

/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
* MPI_SEND_ARRIVED is issued when a MPI message arrives at its destination (the message is transported by model-net and an event is invoked when it arrives. 
* MPI_SEND_POSTED is issued when a MPI message has left the source LP (message is transported via model-net). */
enum MPI_NW_EVENTS
{
	MPI_OP_GET_NEXT=1,
	MPI_SEND_ARRIVED,
    MPI_SEND_ARRIVED_CB, // for tracking message times on sender
	MPI_SEND_POSTED,
};

/* stores pointers of pending MPI operations to be matched with their respective sends/receives. */
struct mpi_msgs_queue
{
    int op_type;
    int tag;
    int source_rank;
    int dest_rank;
    int num_bytes;
    tw_stime req_init_time;
	dumpi_req_id req_id;
    struct qlist_head ql;
};

/* stores request IDs of completed MPI operations (Isends or Irecvs) */
struct completed_requests
{
	dumpi_req_id req_id;
    struct qlist_head ql;
};

/* for wait operations, store the pending operation and number of completed waits so far. */
struct pending_waits
{
    int op_type;
95
    int32_t req_ids[MAX_WAIT_REQS];
96
	int num_completed;
97 98
	int count;
    tw_stime start_time;
99 100 101 102 103 104 105 106 107 108 109 110 111 112 113
    struct qlist_head ql;
};

typedef struct mpi_msgs_queue mpi_msgs_queue;
typedef struct completed_requests completed_requests;
typedef struct pending_waits pending_waits;

/* state of the network LP. It contains the pointers to send/receive lists */
struct nw_state
{
	long num_events_per_lp;
	tw_lpid nw_id;
	short wrkld_end;

    struct rc_stack * processed_ops;
114
    struct rc_stack * matched_reqs;
115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141

    /* count of sends, receives, collectives and delays */
	unsigned long num_sends;
	unsigned long num_recvs;
	unsigned long num_cols;
	unsigned long num_delays;
	unsigned long num_wait;
	unsigned long num_waitall;
	unsigned long num_waitsome;

	/* time spent by the LP in executing the app trace*/
	double start_time;
	double elapsed_time;
	/* time spent in compute operations */
	double compute_time;
	/* time spent in message send/isend */
	double send_time;
	/* time spent in message receive */
	double recv_time;
	/* time spent in wait operation */
	double wait_time;
	/* FIFO for isend messages arrived on destination */
	struct qlist_head arrival_queue;
	/* FIFO for irecv messages posted but not yet matched with send operations */
	struct qlist_head pending_recvs_queue;
	/* List of completed send/receive requests */
	struct qlist_head completed_reqs;
142 143 144

    /* Pending wait operation */
    struct pending_waits * wait_op;
145 146 147 148 149

    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

    char output_buf[512];
150 151 152 153 154 155 156 157
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
* ptr_match_op holds the matched MPI operation which are removed from the queues when a send is matched with the receive in forward event handler. 
* network event being sent. op is the MPI operation issued by the network workloads API. rv_data holds the data for reverse computation (TODO: Fill this data structure only when the simulation runs in optimistic mode). */
struct nw_message
{
158
   // forward message handler 
159
   int msg_type;
160 161
   int op_type;
   
162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183
   struct
   {
       tw_lpid src_rank;
       tw_lpid dest_rank;
       int num_bytes;
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
       int16_t req_id;   
       int tag;
       int found_match;
       short wait_completed;
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
       int saved_num_bytes;
184
       struct codes_workload_op * saved_op;
185
   } rc;
186 187 188 189
};

/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
190
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
191 192
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
193
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
194 195
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
196
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
197 198
/* execute the computational delay */
static void codes_exec_comp_delay(
199
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
200 201 202 203 204 205 206 207 208 209 210 211 212 213 214
/* gets the next MPI operation from the network-workloads API. */
static void get_next_mpi_operation(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
/* reverse handler of get next mpi operation. */
static void get_next_mpi_operation_rc(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
/* Makes a call to get_next_mpi_operation. */
static void codes_issue_next_event(tw_lp* lp);
/* reverse handler of next operation */
static void codes_issue_next_event_rc(tw_lp* lp);


///////////////////// HELPER FUNCTIONS FOR MPI MESSAGE QUEUE HANDLING ///////////////
/* upon arrival of local completion message, inserts operation in completed send queue */
/* upon arrival of an isend operation, updates the arrival queue of the network */
215 216 217 218 219 220 221 222
static void update_completed_queue(
        nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp, dumpi_req_id req_id);
/* reverse of the above function */
static void update_completed_queue_rc(
        nw_state*s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp);
223 224 225 226 227 228 229 230 231 232 233 234 235 236 237
static void update_arrival_queue(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
/* reverse of the above function */
static void update_arrival_queue_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
/* callback to a message sender for computing message time */
static void update_message_time(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
/* reverse for computing message time */
static void update_message_time_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);

/* conversion from seconds to eanaoseconds */
static tw_stime s_to_ns(tw_stime ns);

238
/* Debugging functions, may generate unused function warning */
239
static void print_waiting_reqs(int32_t * reqs, int count)
240 241 242 243
{
    printf("\n Waiting reqs: ");
    int i;
    for(i = 0; i < count; i++ )
244
        printf(" %d ", reqs[i]);
245 246 247 248 249 250 251 252 253
}
static void print_completed_queue(struct qlist_head * head)
{
    printf("\n Completed queue: ");
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
254
            printf(" %d ", current->req_id);
255 256
       }
}
257
static int clear_completed_reqs(nw_state * s,
258
        tw_lp * lp,
259
        int32_t * reqs, int count)
260
{
261
    int i, matched = 0;
262 263 264 265 266 267 268 269 270
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
            struct completed_requests* current = 
                qlist_entry(ent, completed_requests, ql);
            if(current->req_id == reqs[i])
            {
271
                ++matched;
272 273 274 275 276
                qlist_del(&current->ql);
                rc_stack_push(lp, current, free, s->matched_reqs);
            }
       }
    }
277
    return matched;
278 279
}
static void add_completed_reqs(nw_state * s, 
280 281
        tw_lp * lp,
        int count)
282 283 284 285 286 287 288 289
{
    int i;
    for( i = 0; i < count; i++)
    {
       struct completed_requests * req = rc_stack_pop(s->matched_reqs); 
       qlist_add(&req->ql, &s->completed_reqs); 
    }
}
290 291 292 293 294 295
/* helper function - maps an MPI rank to an LP id */
static tw_lpid rank_to_lpid(int rank)
{
    return codes_mapping_get_lpid_from_relative(rank, NULL, "nw-lp", NULL, 0);
}

296 297 298
static int notify_posted_wait(nw_state* s,
        tw_bf * bf, nw_message * m, tw_lp * lp, 
        dumpi_req_id completed_req)
299
{
300 301
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
302

303 304
    m->fwd.wait_completed = 0;
    
305 306
    if(!wait_elem)
        return 0;
307
    
308
    int op_type = wait_elem->op_type;
309

310 311 312 313 314 315 316 317 318 319 320 321 322
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
            wait_completed = 1;
    }
    else if(op_type == CODES_WK_WAITALL 
            || op_type == CODES_WK_WAITANY 
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
323
            {
324
                wait_elem->num_completed++;
325
                if(wait_elem->num_completed > wait_elem->count)
326
                    printf("\n Num completed %d count %d LP %llu ",
327 328 329
                            wait_elem->num_completed,
                            wait_elem->count,
                            lp->gid);
330 331 332 333
                assert(wait_elem->num_completed <= wait_elem->count);
                if(wait_elem->num_completed == wait_elem->count)
                    wait_completed = 1;
           
334
                m->fwd.wait_completed = 1;
335
            }
336
        }
337
    }
338
    return wait_completed;
339
}
340

341
/* reverse handler of MPI wait operation */
342
static void codes_exec_mpi_wait_rc(nw_state* s, tw_lp* lp)
343
{
344
    if(s->wait_op)
345
     {
346 347 348
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
349 350 351 352
     }
   else
    {
        codes_issue_next_event_rc(lp);
353
        completed_requests * qi = rc_stack_pop(s->processed_ops);
354
        qlist_add(&qi->ql, &s->completed_reqs);
355
    }
356
    return;
357
}
358

359
/* execute MPI wait operation */
360
static void codes_exec_mpi_wait(nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
361
{
362 363
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->wait_op);
364
    dumpi_req_id req_id = mpi_op->u.wait.req_id;
365
    struct completed_requests* current = NULL;
366

367 368 369 370 371 372 373
    struct qlist_head * ent = NULL;
    qlist_for_each(ent, &s->completed_reqs)
    {
        current = qlist_entry(ent, completed_requests, ql);
        if(current->req_id == req_id)
        {
            qlist_del(&current->ql);
374
            rc_stack_push(lp, current, free, s->processed_ops);
375 376 377 378
            codes_issue_next_event(lp);
            return;
        }
    }
379 380 381 382 383
    /* If not, add the wait operation in the pending 'waits' list. */
    struct pending_waits* wait_op = malloc(sizeof(struct pending_waits));
    wait_op->op_type = mpi_op->op_type;
    wait_op->req_ids[0] = req_id;
    wait_op->count = 1;
384 385
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
386
    s->wait_op = wait_op;
387

388
    return;
389 390
}

391 392 393 394
static void codes_exec_mpi_wait_all_rc(
        nw_state* s, 
        tw_bf * bf,
        nw_message * m,
395
        tw_lp* lp)
396
{
397 398 399 400 401 402 403 404
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
405
      add_completed_reqs(s, lp, m->fwd.num_matched);
406 407 408
      codes_issue_next_event_rc(lp);
  }
  return;
409 410
}
static void codes_exec_mpi_wait_all(
411 412 413 414 415
        nw_state* s, 
        tw_bf * bf,
        nw_message * m,
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
416 417
{
  int count = mpi_op->u.waits.count;
418 419
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
420

421
  int i = 0, num_matched = 0;
422
  m->fwd.num_matched = 0;
423

424
  /*if(lp->gid == TRACK)
425
  {
426
      printf("\n MPI Wait all posted ");
427 428
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
      print_completed_queue(&s->completed_reqs);
429
  }*/
430 431 432 433 434 435 436 437 438 439 440 441 442
      /* check number of completed irecvs in the completion queue */ 
  for(i = 0; i < count; i++)
  {
      dumpi_req_id req_id = mpi_op->u.waits.req_ids[i];
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
            current = qlist_entry(ent, completed_requests, ql);
            if(current->req_id == req_id)
                num_matched++;
       }
  }
443

444
  m->fwd.found_match = num_matched;
445 446 447 448
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
449 450 451
      m->fwd.num_matched = clear_completed_reqs(s, lp, mpi_op->u.waits.req_ids, count);
      struct pending_waits* wait_op = s->wait_op;
      free(wait_op);
452 453
      s->wait_op = NULL;
      codes_issue_next_event(lp);
454 455
  }
  else
456 457 458 459 460 461 462 463 464 465 466
  {
      /* If not, add the wait operation in the pending 'waits' list. */
	  struct pending_waits* wait_op = malloc(sizeof(struct pending_waits));
	  wait_op->count = count;
      wait_op->op_type = mpi_op->op_type;
      assert(count < MAX_WAIT_REQS);

      for(i = 0; i < count; i++)
          wait_op->req_ids[i] =  mpi_op->u.waits.req_ids[i];

	  wait_op->num_completed = num_matched;
467
	  wait_op->start_time = tw_now(lp);
468
      s->wait_op = wait_op;
469
  }
470 471
  return;
}
472 473 474 475

/* search for a matching mpi operation and remove it from the list. 
 * Record the index in the list from where the element got deleted. 
 * Index is used for inserting the element once again in the queue for reverse computation. */
476 477 478 479 480
static int rm_matching_rcv(nw_state * ns, 
        tw_bf * bf,
        nw_message * m, 
        tw_lp * lp, 
        mpi_msgs_queue * qitem)
481 482
{
    int matched = 0;
483
    int index = 0;
484 485
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
486

487 488
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
489
        if((qi->num_bytes == qitem->num_bytes)
490 491
                && ((qi->tag == qitem->tag) || qi->tag == -1)
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
492 493 494 495
        {
            matched = 1;
            break;
        }
496
        ++index;
497 498 499 500
    }
    
    if(matched)
    {
501
        m->rc.saved_recv_time = ns->recv_time;
502
        ns->recv_time += (tw_now(lp) - qi->req_init_time);
503 504 505
        
        if(qi->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qi->req_id);
506
        
507
        qlist_del(&qi->ql);
508 509
        
        rc_stack_push(lp, qi, free, ns->processed_ops);
510
        return index;
511 512 513 514
    }
    return -1;
}

515 516 517 518
static int rm_matching_send(nw_state * ns, 
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
519 520 521 522 523
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

524
    int index = 0;
525 526
    qlist_for_each(ent, &ns->arrival_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
527
        if((qi->num_bytes == qitem->num_bytes) 
528 529 530 531 532 533
                && (qi->tag == qitem->tag || qitem->tag == -1)
                && ((qi->source_rank == qitem->source_rank) || qitem->source_rank == -1))
        {
            matched = 1;
            break;
        }
534
        ++index;
535 536 537 538
    }

    if(matched)
    {
539
        m->rc.saved_recv_time = ns->recv_time;
540 541 542 543 544
        ns->recv_time += (tw_now(lp) - qitem->req_init_time);

        if(qitem->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qitem->req_id);

545
        qlist_del(&qi->ql);
546

547
        return index;
548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573
    }
    return -1;
}
static void codes_issue_next_event_rc(tw_lp * lp)
{
	    tw_rand_reverse_unif(lp->rng);	
}

/* Trigger getting next event at LP */
static void codes_issue_next_event(tw_lp* lp)
{
   tw_event *e;
   nw_message* msg;

   tw_stime ts;

   ts = g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, noise);
   e = tw_event_new( lp->gid, ts, lp );
   msg = tw_event_data(e);

   msg->msg_type = MPI_OP_GET_NEXT;
   tw_event_send(e);
}

/* Simulate delays between MPI operations */
static void codes_exec_comp_delay(
574
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
575 576 577 578 579
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

580
    m->rc.saved_delay = s->compute_time;
581 582
    s->compute_time += s_to_ns(mpi_op->u.delay.seconds);
    ts = s_to_ns(mpi_op->u.delay.seconds);
583 584 585 586 587 588 589 590 591 592 593

	ts += g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, noise);
	
	e = tw_event_new( lp->gid, ts , lp );
	msg = tw_event_data(e);
	msg->msg_type = MPI_OP_GET_NEXT;
	tw_event_send(e); 
                
}

/* reverse computation operation for MPI irecv */
594 595 596 597
static void codes_exec_mpi_recv_rc(
        nw_state* ns, 
        tw_bf * bf, 
        nw_message* m, 
598
        tw_lp* lp)
599
{
600
	num_bytes_recvd -= m->rc.saved_num_bytes;
601 602
	ns->recv_time = m->rc.saved_recv_time;
	if(m->fwd.found_match >= 0)
603
	  {
604
		ns->recv_time = m->rc.saved_recv_time;
605
        int queue_count = qlist_count(&ns->arrival_queue); 
606
        
607
        mpi_msgs_queue * qi = rc_stack_pop(ns->processed_ops);	
608
       
609
        if(!m->fwd.found_match)
610 611 612
        {
            qlist_add(&qi->ql, &ns->arrival_queue);
        }
613
        else if(m->fwd.found_match >= queue_count)
614 615 616
        {
            qlist_add_tail(&qi->ql, &ns->arrival_queue);
        }
617
        else if(m->fwd.found_match > 0 && m->fwd.found_match < queue_count) 
618
        {
619 620 621 622
            int index = 1;
            struct qlist_head * ent = NULL;
            qlist_for_each(ent, &ns->arrival_queue)
            {
623
               if(index == m->fwd.found_match)
624 625 626 627 628 629
               {
                 qlist_add(&qi->ql, ent);
                 break;
               }
               index++; 
            }
630
        }
631
        if(qi->op_type == CODES_WK_IRECV)
632
        {
633
            update_completed_queue_rc(ns, bf, m, lp);
634
        }
635 636
        codes_issue_next_event_rc(lp);
      }
637
	else if(m->fwd.found_match < 0)
638
	    {
639 640 641 642
            struct qlist_head * ent = qlist_pop_back(&ns->pending_recvs_queue); 
            mpi_msgs_queue * qi = qlist_entry(ent, mpi_msgs_queue, ql);
            free(qi);
            
643
            if(m->op_type == CODES_WK_IRECV)
644
                codes_issue_next_event_rc(lp);
645 646 647 648
	    }
}

/* Execute MPI Irecv operation (non-blocking receive) */ 
649 650 651 652 653 654
static void codes_exec_mpi_recv(
        nw_state* s, 
        tw_bf * bf,
        nw_message * m, 
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
655 656 657 658 659
{
/* Once an irecv is posted, list of completed sends is checked to find a matching isend.
   If no matching isend is found, the receive operation is queued in the pending queue of
   receive operations. */

660
	m->rc.saved_recv_time = s->recv_time;
661 662
    m->rc.saved_num_bytes = mpi_op->u.recv.num_bytes;

663 664 665 666 667 668 669 670 671 672 673
	num_bytes_recvd += mpi_op->u.recv.num_bytes;

    mpi_msgs_queue * recv_op = (mpi_msgs_queue*) malloc(sizeof(mpi_msgs_queue));
    recv_op->req_init_time = tw_now(lp);
    recv_op->op_type = mpi_op->op_type;
    recv_op->source_rank = mpi_op->u.recv.source_rank;
    recv_op->dest_rank = mpi_op->u.recv.dest_rank;
    recv_op->num_bytes = mpi_op->u.recv.num_bytes;
    recv_op->tag = mpi_op->u.recv.tag;
    recv_op->req_id = mpi_op->u.recv.req_id;

674 675 676 677
    if(s->nw_id == TRACK_LP)
        printf("\n Receive op posted num bytes %d source %d ", recv_op->num_bytes,
                recv_op->source_rank);

678
	int found_matching_sends = rm_matching_send(s, bf, m, lp, recv_op);
679 680 681 682

	/* save the req id inserted in the completed queue for reverse computation. */
	if(found_matching_sends < 0)
	  {
683
	   	  m->fwd.found_match = -1;
684
          qlist_add_tail(&recv_op->ql, &s->pending_recvs_queue);
685 686 687 688 689 690 691
	
	       /* for mpi irecvs, this is a non-blocking receive so just post it and move on with the trace read. */
		if(mpi_op->op_type == CODES_WK_IRECV)
		   {
			codes_issue_next_event(lp);	
			return;
		   }
692
      }
693 694
	else
	  {
695
        m->fwd.found_match = found_matching_sends;
696
        codes_issue_next_event(lp); 
697 698
	    rc_stack_push(lp, recv_op, free, s->processed_ops);
      }
699 700 701
}

/* executes MPI send and isend operations */
702 703 704 705 706
static void codes_exec_mpi_send(nw_state* s, 
        tw_bf * bf,
        nw_message * m,
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
707
{
708
    m->rc.saved_num_bytes = mpi_op->u.send.num_bytes;
709 710 711 712 713 714 715 716
	/* model-net event */
	tw_lpid dest_rank;
	codes_mapping_get_lp_info(lp->gid, lp_group_name, &mapping_grp_id, 
	    lp_type_name, &mapping_type_id, annotation, &mapping_rep_id, &mapping_offset);

	if(net_id == DRAGONFLY) /* special handling for the dragonfly case */
	{
		int num_routers, lps_per_rep, factor;
717 718
		num_routers = codes_mapping_get_lp_count(lp_group_name, 1,
                  "modelnet_dragonfly_router", NULL, 1);
719 720 721 722 723 724 725 726 727 728 729 730
	 	lps_per_rep = (2 * num_nw_lps) + num_routers;	
		factor = mpi_op->u.send.dest_rank / num_nw_lps;
		dest_rank = (lps_per_rep * factor) + (mpi_op->u.send.dest_rank % num_nw_lps);	
	}
	else
	{
		/* other cases like torus/simplenet/loggp etc. */
		codes_mapping_get_lp_id(lp_group_name, lp_type_name, NULL, 1,  
	    	  mpi_op->u.send.dest_rank, mapping_offset, &dest_rank);
	}

	num_bytes_sent += mpi_op->u.send.num_bytes;
731
    s->num_bytes_sent += mpi_op->u.send.num_bytes;
732 733 734 735

	nw_message local_m;
	nw_message remote_m;

736 737 738
    local_m.fwd.sim_start_time = tw_now(lp);
    local_m.fwd.dest_rank = mpi_op->u.send.dest_rank;
    local_m.fwd.src_rank = mpi_op->u.send.source_rank;
739
    local_m.op_type = mpi_op->op_type; 
740
    local_m.msg_type = MPI_SEND_POSTED;
741 742 743
    local_m.fwd.tag = mpi_op->u.send.tag;
    local_m.fwd.num_bytes = mpi_op->u.send.num_bytes;
    local_m.fwd.req_id = mpi_op->u.send.req_id;
744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761

    remote_m = local_m;
	remote_m.msg_type = MPI_SEND_ARRIVED;

	model_net_event(net_id, "test", dest_rank, mpi_op->u.send.num_bytes, 0.0, 
	    sizeof(nw_message), (const void*)&remote_m, sizeof(nw_message), (const void*)&local_m, lp);

	/* isend executed, now get next MPI operation from the queue */ 
	if(mpi_op->op_type == CODES_WK_ISEND)
	   codes_issue_next_event(lp);
}

/* convert seconds to ns */
static tw_stime s_to_ns(tw_stime ns)
{
    return(ns * (1000.0 * 1000.0 * 1000.0));
}

762 763
static void update_completed_queue_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
764
   
765 766 767
    if(bf->c0)
    {
       struct qlist_head * ent = qlist_pop_back(&s->completed_reqs);
768 769 770 771 772 773 774

        completed_requests * req = qlist_entry(ent, completed_requests, ql);
      /*if(lp->gid == TRACK)
      {
          printf("\n After popping %ld ", req->req_id);
        print_completed_queue(&s->completed_reqs);
      }*/
775 776 777 778
       free(req);
    }
    else if(bf->c1)
    {
779
       struct pending_waits* wait_elem = rc_stack_pop(s->processed_ops); 
780
       s->wait_op = wait_elem;
781 782
       s->wait_time = m->rc.saved_wait_time;
       add_completed_reqs(s, lp, m->fwd.num_matched);
783 784
       codes_issue_next_event_rc(lp); 
    }
785 786
    if(m->fwd.wait_completed > 0)
           s->wait_op->num_completed--;
787 788 789 790 791 792 793 794 795 796
}

static void update_completed_queue(nw_state* s, 
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp,
        dumpi_req_id req_id)
{
    bf->c0 = 0;
    bf->c1 = 0;
797
    m->fwd.num_matched = 0; 
798

799 800
    int waiting = 0;
    waiting = notify_posted_wait(s, bf, m, lp, req_id);
801 802 803 804 805 806 807
 
    if(!waiting)
    {
        bf->c0 = 1;
        completed_requests * req = malloc(sizeof(completed_requests));
        req->req_id = req_id;
        qlist_add_tail(&req->ql, &s->completed_reqs);
808 809 810 811 812 813
       
        /*if(lp->gid == TRACK)
        {
            printf("\n Forward mode adding %ld ", req_id);
            print_completed_queue(&s->completed_reqs);
        }*/
814 815 816 817
    }
    else 
     {
            bf->c1 = 1;
818 819
            m->fwd.num_matched = clear_completed_reqs(s, lp, s->wait_op->req_ids, s->wait_op->count);
            m->rc.saved_wait_time = s->wait_time;
820
            s->wait_time += (tw_now(lp) - s->wait_op->start_time);  
821 822 823

            struct pending_waits* wait_elem = s->wait_op;
            rc_stack_push(lp, wait_elem, free, s->processed_ops);
824 825 826 827 828
            s->wait_op = NULL;
            codes_issue_next_event(lp); 
     }
}

829
/* reverse handler for updating arrival queue function */
830 831 832
static void update_arrival_queue_rc(nw_state* s, 
        tw_bf * bf, 
        nw_message * m, tw_lp * lp)
833
{
834
	s->recv_time = m->rc.saved_recv_time;
835 836
    s->num_bytes_recvd -= m->fwd.num_bytes;

837
    codes_local_latency_reverse(lp);
838
  
839
    if(m->fwd.found_match >= 0)
840
	{
841
        mpi_msgs_queue * qi = rc_stack_pop(s->processed_ops);
842 843
        int queue_count = qlist_count(&s->pending_recvs_queue); 

844
        if(!m->fwd.found_match)
845 846 847
        {
            qlist_add(&qi->ql, &s->pending_recvs_queue);
        }
848
        else if(m->fwd.found_match >= queue_count)
849
        {
850 851
            qlist_add_tail(&qi->ql, &s->pending_recvs_queue);
        }
852
        else if(m->fwd.found_match > 0 && m->fwd.found_match < queue_count)
853 854 855 856 857
        {
            int index = 1;
            struct qlist_head * ent = NULL;
            qlist_for_each(ent, &s->pending_recvs_queue)
            {
858
               if(index == m->fwd.found_match)
859 860 861 862 863 864
               {
                 qlist_add(&qi->ql, ent);
                 break;
               }
               index++; 
            }
865
        }
866 867
        if(qi->op_type == CODES_WK_IRECV)
            update_completed_queue_rc(s, bf, m, lp);
868
    }
869
	else if(m->fwd.found_match < 0)
870 871 872 873 874 875 876 877 878 879
	{
	    struct qlist_head * ent = qlist_pop_back(&s->arrival_queue); 
        mpi_msgs_queue * qi = qlist_entry(ent, mpi_msgs_queue, ql);
        free(qi);
    }
}

/* once an isend operation arrives, the pending receives queue is checked to find out if there is a irecv that has already been posted. If no isend has been posted, */
static void update_arrival_queue(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
880
	m->rc.saved_recv_time = s->recv_time;
881
    s->num_bytes_recvd += m->fwd.num_bytes;
882 883 884

    // send a callback to the sender to increment times
    tw_event *e_callback =
885
        tw_event_new(rank_to_lpid(m->fwd.src_rank),
886 887 888
                codes_local_latency(lp), lp);
    nw_message *m_callback = tw_event_data(e_callback);
    m_callback->msg_type = MPI_SEND_ARRIVED_CB;
889
    m_callback->fwd.msg_send_time = tw_now(lp) - m->fwd.sim_start_time;
890 891 892 893
    tw_event_send(e_callback);

    /* Now reconstruct the queue item */
    mpi_msgs_queue * arrived_op = (mpi_msgs_queue *) malloc(sizeof(mpi_msgs_queue));
894
    arrived_op->req_init_time = m->fwd.sim_start_time;
895
    arrived_op->op_type = m->op_type;
896 897 898 899
    arrived_op->source_rank = m->fwd.src_rank;
    arrived_op->dest_rank = m->fwd.dest_rank;
    arrived_op->num_bytes = m->fwd.num_bytes;
    arrived_op->tag = m->fwd.tag;
900

901
    if(s->nw_id == TRACK_LP)
902
        printf("\n Send op arrived source rank %d num bytes %d ", arrived_op->source_rank,
903 904
                arrived_op->num_bytes);

905
    int found_matching_recv = rm_matching_rcv(s, bf, m, lp, arrived_op);
906 907 908

    if(found_matching_recv < 0)
    {
909
        m->fwd.found_match = -1;
910 911 912 913
        qlist_add_tail(&arrived_op->ql, &s->arrival_queue);
    }
    else
    {
914
        m->fwd.found_match = found_matching_recv;
915
        free(arrived_op);
916 917 918 919 920 921 922 923
    }
}
static void update_message_time(
        nw_state * s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp)
{
924 925
    m->rc.saved_send_time = s->send_time;
    s->send_time += m->fwd.msg_send_time;
926 927 928 929 930 931 932 933
}

static void update_message_time_rc(
        nw_state * s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp)
{
934
    s->send_time = m->rc.saved_send_time;
935 936 937 938 939 940 941 942
}

/* initializes the network node LP, loads the trace file in the structs, calls the first MPI operation to be executed */
void nw_test_init(nw_state* s, tw_lp* lp)
{
   /* initialize the LP's and load the data */
   char * params = NULL;
   dumpi_trace_params params_d;
943
 
944
   memset(s, 0, sizeof(*s));
945
   s->nw_id = codes_mapping_get_lp_relative_id(lp->gid, 0, 0);
946 947 948 949

   if(!num_net_traces) 
	num_net_traces = num_net_lps;

950 951
   assert(num_net_traces <= num_net_lps);

952 953 954 955 956 957 958 959 960 961 962 963
   if (strcmp(workload_type, "dumpi") == 0){
       strcpy(params_d.file_name, workload_file);
       params_d.num_net_traces = num_net_traces;

       params = (char*)&params_d;
   }
  /* In this case, the LP will not generate any workload related events*/
   if(s->nw_id >= params_d.num_net_traces)
	    return;

   /* Initialize the RC stack */
   rc_stack_create(&s->processed_ops);
964
   rc_stack_create(&s->matched_reqs);
965 966

   assert(s->processed_ops != NULL);
967
   assert(s->matched_reqs != NULL);
968 969 970 971 972 973 974 975 976 977

   wrkld_id = codes_workload_load("dumpi-trace-workload", params, 0, (int)s->nw_id);

   INIT_QLIST_HEAD(&s->arrival_queue);
   INIT_QLIST_HEAD(&s->pending_recvs_queue);
   INIT_QLIST_HEAD(&s->completed_reqs);

   /* clock starts when the first event is processed */
   s->start_time = tw_now(lp);
   codes_issue_next_event(lp);
978 979
   s->num_bytes_sent = 0;
   s->num_bytes_recvd = 0;
980 981
   s->compute_time = 0;
   s->elapsed_time = 0;
982 983 984 985 986 987 988

   return;
}

void nw_test_event_handler(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
	*(int *)bf = (int)0;
989 990
    rc_stack_gc(lp, s->matched_reqs);
    rc_stack_gc(lp, s->processed_ops);
991 992 993 994 995 996 997 998 999 1000 1001

    switch(m->msg_type)
	{
		case MPI_SEND_ARRIVED:
			update_arrival_queue(s, bf, m, lp);
		break;

		case MPI_SEND_ARRIVED_CB:
			update_message_time(s, bf, m, lp);
		break;

1002 1003 1004 1005 1006 1007 1008
        case MPI_SEND_POSTED:
        {
           if(m->op_type == CODES_WK_SEND)
               codes_issue_next_event(lp);
           else
            if(m->op_type == CODES_WK_ISEND)
            {
1009
              update_completed_queue(s, bf, m, lp, m->fwd.req_id);  
1010 1011 1012
            }
        }
        break;
1013 1014 1015 1016 1017 1018 1019 1020
		case MPI_OP_GET_NEXT:
			get_next_mpi_operation(s, bf, m, lp);	
		break; 
	}
}

static void get_next_mpi_operation_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
1021
    codes_workload_get_next_rc2(wrkld_id, 0, (int)s->nw_id);
1022

1023
	if(m->op_type == CODES_WK_END)
1024
    {
1025
		return;
1026
    }
1027
	switch(m->op_type)
1028 1029 1030 1031
	{
		case CODES_WK_SEND:
		case CODES_WK_ISEND:
		{
1032
			model_net_event_rc(net_id, lp, m->rc.saved_num_bytes);
1033
			if(m->op_type == CODES_WK_ISEND)
1034 1035
				codes_issue_next_event_rc(lp);
			s->num_sends--;
1036 1037
            s->num_bytes_sent += m->rc.saved_num_bytes;
			num_bytes_sent -= m->rc.saved_num_bytes;
1038 1039 1040 1041 1042 1043
		}
		break;

		case CODES_WK_IRECV:
		case CODES_WK_RECV:
		{
1044
			codes_exec_mpi_recv_rc(s, bf, m, lp);
1045 1046 1047
			s->num_recvs--;
		}
		break;
1048 1049 1050
        
		
        case CODES_WK_DELAY:
1051 1052
		{
			s->num_delays--;
1053
            tw_rand_reverse_unif(lp->rng);
1054 1055
            if(!disable_delay)
                s->compute_time = m->rc.saved_delay;
1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071
		}
		break;
		case CODES_WK_BCAST:
		case CODES_WK_ALLGATHER:
		case CODES_WK_ALLGATHERV:
		case CODES_WK_ALLTOALL:
		case CODES_WK_ALLTOALLV:
		case CODES_WK_REDUCE:
		case CODES_WK_ALLREDUCE:
		case CODES_WK_COL:
		{
			s->num_cols--;
		    codes_issue_next_event_rc(lp);
        }
		break;
	
1072 1073 1074 1075 1076 1077 1078 1079
		case CODES_WK_WAITSOME:
		case CODES_WK_WAITANY:
        {
           s->num_waitsome--;
           codes_issue_next_event_rc(lp); 
        }
        break;

1080 1081 1082
		case CODES_WK_WAIT:
		{
			s->num_wait--;
1083
			codes_exec_mpi_wait_rc(s, lp);
1084 1085 1086 1087 1088
		}
		break;
		case CODES_WK_WAITALL:
		{
			s->num_waitall--;
1089
            codes_exec_mpi_wait_all_rc(s, bf, m, lp);
1090 1091 1092
		}
		break;
		default:
1093
			printf("\n Invalid op type %d ", m->op_type);
1094 1095 1096 1097 1098
	}
}

static void get_next_mpi_operation(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
1099 1100 1101
		//struct codes_workload_op * mpi_op = malloc(sizeof(struct codes_workload_op));
        struct codes_workload_op mpi_op;
        codes_workload_get_next(wrkld_id, 0, (int)s->nw_id, &mpi_op);
1102
      
1103
        m->op_type = mpi_op.op_type;
1104

1105
        if(mpi_op.op_type == CODES_WK_END)
1106 1107 1108 1109
        {
            s->elapsed_time = tw_now(lp) - s->start_time;
            return;
        }
1110
		switch(mpi_op.op_type)
1111 1112 1113 1114 1115
		{
			case CODES_WK_SEND:
			case CODES_WK_ISEND:
			 {
				s->num_sends++;
1116
				codes_exec_mpi_send(s, bf, m, lp, &mpi_op);
1117 1118 1119 1120 1121 1122 1123
			 }
			break;
	
			case CODES_WK_RECV:
			case CODES_WK_IRECV:
			  {
				s->num_recvs++;
1124
				codes_exec_mpi_recv(s, bf, m, lp, &mpi_op);
1125 1126 1127 1128 1129
			  }
			break;

			case CODES_WK_DELAY:
			  {
1130
                
1131
				s->num_delays++;
1132 1133 1134
                if(disable_delay)
                    codes_issue_next_event(lp);
                else
1135
				    codes_exec_comp_delay(s, m, lp, &mpi_op);
1136 1137 1138 1139 1140
			  }
			break;

            case CODES_WK_WAITSOME:
            case CODES_WK_WAITANY:
1141 1142 1143 1144 1145 1146
            {
                s->num_waitsome++;
                codes_issue_next_event(lp);
            }
            break;

1147
			case CODES_WK_WAITALL:
1148
			  {
1149
				s->num_waitall++;
1150
			    codes_exec_mpi_wait_all(s, bf, m, lp, &mpi_op);
1151
              }
1152 1153 1154 1155
			break;
			case CODES_WK_WAIT:
			{
				s->num_wait++;
1156
                codes_exec_mpi_wait(s, lp, &mpi_op);
1157 1158
			}
			break;
1159 1160 1161 1162 1163 1164 1165 1166
			case CODES_WK_BCAST:
			case CODES_WK_ALLGATHER:
			case CODES_WK_ALLGATHERV:
			case CODES_WK_ALLTOALL:
			case CODES_WK_ALLTOALLV:
			case CODES_WK_REDUCE:
			case CODES_WK_ALLREDUCE:
			case CODES_WK_COL:
1167
			{
1168
				s->num_cols++;
1169 1170 1171 1172
			    codes_issue_next_event(lp);
            }
			break;
			default:
1173
				printf("\n Invalid op type %d ", mpi_op.op_type);
1174 1175 1176 1177 1178 1179
		}
        return;
}

void nw_test_finalize(nw_state* s, tw_lp* lp)
{
1180 1181 1182
    int written = 0;
    if(!s->nw_id)
        written = sprintf(s->output_buf, "# Format <LP ID> <Terminal ID> <Total sends> <Total Recvs> <Bytes sent> <Bytes recvd> <Send time> <Comm. time> <Compute time>");
1183
	if(s->nw_id < (tw_lpid)num_net_traces)
1184 1185 1186
	{
		int count_irecv = qlist_count(&s->pending_recvs_queue);
        int count_isend = qlist_count(&s->arrival_queue);
1187
# if DBG_MPI_SIM == 1
1188
		printf("\n LP %llu unmatched irecvs %d unmatched sends %d Total sends %ld receives %ld collectives %ld delays %ld wait alls %ld waits %ld send time %lf wait %lf", 
1189
			lp->gid, count_irecv, count_isend, s->num_sends, s->num_recvs, s->num_cols, s->num_delays, s->num_waitall, s->num_wait, s->send_time, s->wait_time);
1190
# endif
1191
        written += sprintf(s->output_buf + written, "\n %llu %llu %ld %ld %ld %ld %lf %lf %lf", lp->gid, s->nw_id, s->num_sends, s->num_recvs, s->num_bytes_sent, 
1192 1193 1194
                s->num_bytes_recvd, s->send_time, s->elapsed_time - s->compute_time, s->compute_time);
        lp_io_write(lp->gid, "mpi-replay-stats", written, s->output_buf);

1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216
		if(s->elapsed_time - s->compute_time > max_comm_time)
			max_comm_time = s->elapsed_time - s->compute_time;
		
		if(s->elapsed_time > max_time )
			max_time = s->elapsed_time;

		/*if(s->wait_time > max_wait_time)
			max_wait_time = s->wait_time;
        */
		if(s->send_time > max_send_time)
			max_send_time = s->send_time;

		if(s->recv_time > max_recv_time)
			max_recv_time = s->recv_time;

		avg_time += s->elapsed_time;
		avg_comm_time += (s->elapsed_time - s->compute_time);
		avg_wait_time += s->wait_time;
		avg_send_time += s->send_time;
		 avg_recv_time += s->recv_time;

		//printf("\n LP %ld Time spent in communication %llu ", lp->gid, total_time - s->compute_time);
1217
	    rc_stack_destroy(s->matched_reqs);    
1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233
	    rc_stack_destroy(s->processed_ops);    
    }
}

void nw_test_event_handler_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
	switch(m->msg_type)
	{
		case MPI_SEND_ARRIVED:
			update_arrival_queue_rc(s, bf, m, lp);
		break;

		case MPI_SEND_ARRIVED_CB:
			update_message_time_rc(s, bf, m, lp);
		break;

1234 1235 1236 1237 1238 1239 1240 1241 1242 1243
        case MPI_SEND_POSTED:
        {
         if(m->op_type == CODES_WK_SEND) 
             codes_issue_next_event_rc(lp);
         else if(m->op_type == CODES_WK_ISEND)
            update_completed_queue_rc(s, bf, m, lp);  
        }
        break;
		
        case MPI_OP_GET_NEXT:
1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264